MiniMax H3 VRAM requirements and real render times, consolidated from 20 threads

The MiniMax H3 VRAM requirements and render times have been consolidated by a reporter who analyzed 20 threads with contradictory information. The key findings are that the spread of render times on a single card model is larger than the spread between different card models. Factors that affect render times include optimization settings, not hardware capabilities.

Reported render times vary widely based on hardware settings, such as resolution, frame count, step count, and whether Turbo LoRA is enabled. VRAM requirements can be lower than assumed, with reports of successful generation on 12GB VRAM systems with 32 to 64 GB system RAM. Using a low-VRAM workflow can reduce the VRAM floor. A 30-second generation can be achieved by stitching three 10-second clips using comfyui-h3-multishot, which handles seams but is not native.

Local generation using the API is free per clip, but hardware cost is significant upfront. The break-even point for local generation depends on volume and can be calculated based on hardware cost and API rate per second. Local wins on high volume, iteration, custom LoRAs, and not caring about wait times. Hosted solutions are better for low or bursty volume, needing immediate output, lacking suitable hardware, or using 2K resolution.

Most users in the threads are in the local generation group, while those asking about specific hardware compatibility are in a different group. Posting detailed timings with resolution, frames, steps, and optimization settings would be helpful to compare data effectively.